Sizing pits in heat exchanger tubes with eddy current measurements

The project

Heat exchangers are systems to transfer heat for both heating or cooling. They are are used in a wide range of industries, including fridges, cars and power stations. They operate under stress and are therefore prone to developing faults, which can account for significant costs.

One challenge is the characterisation of flaws from Microbial-Induced Corrosion (MIC). Existing methods are able to detect the presence of such flaws but aren't able to reliably predict their depth. Sizing incorrectly may lead to unnecessary removal from service, which affects efficiency, or leaving tubes in services that should be removed, which can lead to unexpected shutdowns.

We set out to develop machine learning models that would reliably predict depth of flaws, to enable better decision-making regarding flaw depth prediciton.

This project was designed by, sponsored by, and undertaken in collaboration with, the Electric Power Research Institute (EPRI).

Team: Chris Lovejoy, Andrea Grafmueller, Chirag Garg, Theodore Hermann

General info
File Structure
Setup
Usage
References

General info

Problem statement

Heat exchanger tubes are commonly inspected with eddy current (EC) techniques. While detection of defects is not a challenge, characterization (that is, determining the depth of wall loss) is done based on a single parameter (amplitude or phase); this mono-parametric approach lacks accuracy and leads to tubes being unnecessarily removed from service. The main defect of interest here is microbiologically-influenced corrosion (MIC). Since this is not reproducible in manufactured tubes, we will start with inner diameter (ID) pits.

Objectives

Primary goal is to develop machine learning models to estimate the depth of defects based on the EC waveforms that perform better than ~15% RMSE (root-mean-squared-error) of the real depth.

Secondary goals pertain to other interesting characterizations (such as shape classification, length and width characterization, etc) but to be only pursued if:

They support the primary goal; or
Primary goal has already been satisfactorily achieved

Outcome

We tested our models on a hold-out test set, and the top performing models achieved an RMSE of approximately 7%.

File-Structure

│   .gitignore
│   README.md
│
├───data
│   │
│   ├─── complete_raw     
│   │        "holds dataframes of the complete original dataset and different subsets"
│   │       full_tube_and_channel_data.csv
│   │       filtered_data.csv
│   │       raw_data.csv
│   │       filtered_cg_data.csv
│   │
│   ├───filtered
│   │       "original filtered data from EPRI"
│   │       [TubeID_SN_FlawID_Angle_filt.csv] x multiple  
│   │       
│   ├───filtered_CG
│   │       "processed data subtracting background differently and fitting peaks"
│   │
│   ├───interim
│   │       "dataframes that hold a collection of the colected features
│   │       with one row per original measurement"
│   │
│   └───raw
│           "original raw data from EPRI"
│           [TubeID_SN_FlawID_Angle_raw.csv] x multiple
│
├───demos
│       "notebooks demonstrating how different functions can be used"
│
├───models
│
├───notebooks
│
└───src
    │
    ├───data
    ├───features
    ├───models
    └───visualisation

Setup

The code was developed using an anaconda virtual environment. To create the environment, please install Anaconda (anaconda.org) and in a command line interface run:

conda env create --file environment-py37.yml

If successful, the command returns instructions on how to activate the environment, for example: conda activate conda-env3_7.

The input data are separate .csv files, each corresponding to one defect, in folder data/raw. From these the code will generate the following interim data:

A single .csv file holding a compilation of all defects of interest created with the CompleteData() class. e.g. filtered_data.csv, stored in complete_raw folder.
A processed .csv file in the interim folder, summarizing the features of each defect measurement in one line, produced with the FeatureData() class. The file contains the following columns: Tube_Alias Flaw_ID Angle Feature_1 .... Feature_N Flaw_Depth Pct_Depth Flaw_Volume Flaw_Area

Note: Notebook 1 contains the code required to do this

Usage

The codebase is organized into custom-build functions located in the src directory and Jupyter notebooks in the notebooks directory. The notebooks show the results and rely on the functions from the src directory.

Note: the data used to train our model is not available, and remains the intellectual property of the Electric Power Research Institute (EPRI).

References

[1] Background info on eddy current measurements.

chris-lovejoy/PitDepthPrediction